dsh-shengtu
Multi-engine image generation for DeepSeek Harness — 生图 / AI image generation, inside the conversation.
It does two things:
- Registers an agent tool
generate_image — say "draw a picture of…" in a conversation and the generated image appears right in the transcript.
- Adds an Image generation button below the composer. It opens a panel with Generate and Config tabs plus an on/off switch.
Install
From the plugin market once it is listed there, or straight from this repository:
dsh plugin add github:xuhairui-dzz/dsh-shengtu
Then reload the page (Ctrl+R). The package declares dsh.bundle, so it activates as a normal bundle in your profile.
To pin a revision instead of tracking main, append the commit:
dsh plugin add github:xuhairui-dzz/dsh-shengtu#<40-char-sha>
Engines (panel → Config → Engine)
| Engine | Credential variable | Default endpoint / model |
|---|
| Doubao · Volcengine Ark | ARK_API_KEY | https://ark.cn-beijing.volces.com/api/v3 / doubao-seedream-5-0-flash-260915 |
| SiliconFlow | GUIJI_API_KEY | https://api.siliconflow.cn/v1 / Kwai-Kolors/Kolors |
| OpenAI | OPENAI_API_KEY | https://api.openai.com/v1 / gpt-image-1 (template) |
| Custom | any | anything exposing POST {baseURL}/images/generations |
Switching engines takes effect immediately; the panel and the agent tool share one configuration and one credential.
Configuration
| Key | Meaning |
|---|
engine | Preset: ark / siliconflow / openai / custom |
baseURL, model | Endpoint and model name |
credentialRef | Credential variable name, defaulted per engine |
size, watermark, timeoutMs | Size, watermark, timeout |
outDir | Output directory; empty = generated/ under the session working directory |
pricePerImage | Unit price (CNY per image) for local spend tracking; 0 counts images only |
Secrets are never written into the config file: they are read through the Harness credential service (settings credential store, shell environment, or .env). The panel's API Key field stores straight into the credential store.
Spend tracking
The panel header shows N today · M total (money only when a unit price is set). This is a local tally, not a provider bill — the real charge is whatever your console says.
Safety boundaries
- Keys are read on the Host side only and are never sent to the browser; the client sees only "ready / missing".
- Downloaded images are validated against a byte cap and format signature (PNG/JPEG/WebP) before being written.
- The panel's image route only serves image files inside
outDir and rejects path traversal.
Development
node test/toggle-test.mjs # offline behavioural test of the on/off switch
License
MIT
